JPH0568716B2 - - Google Patents

Description

[Detailed Description of the Invention] (Technical Field) The present invention relates to a speech recognition method, and more specifically, the present invention relates to a speech recognition method, and specifically, a speech recognition method in which a recognition operation is started from detecting the beginning of input speech without waiting for confirmation of the end of word input speech. Regarding recognition methods.

(Background Art) As a form of speech recognition method, a standard pattern is stored as a frame time series of frequency components corresponding to each standard voice, and the input pattern is extracted from the input speech as a frame time series of frequency components. A known method is to calculate the degree of dissimilarity between an input pattern and each standard pattern, and to identify input speech based on the degree of dissimilarity.

For example, Oki Research and Development No. 118, page 53, Showa
It was disclosed in December 1957. In such a method, the standard or input speech pattern is usually frequency-analyzed by setting frames regularly, and then subjected to logarithmic transformation and normalization of vocal fold vibration characteristics using a least squares approximation straight line, and then the frequency components are determined. , expressed as a frame time series.

In addition, methods for setting matching paths to calculate the degree of dissimilarity between the input pattern and each standard pattern include the DP matching method using dynamic programming and the essentially linear matching method as seen in the above-mentioned literature. The law is known. From the perspective of simplifying the configuration, the linear matching method is advantageous, but
As illustrated in Figure 1, the utterance speed of a word varies extremely widely, with individual differences as well as changes depending on the psychological state and situation. There is some variation in the results, and some kind of improvement is required. Various types of linear matching methods have been proposed, but not only in the above-mentioned literature, in which a matching path is set after detecting the end of input speech, there are problems in terms of recognition response, and there are It is also necessary to memorize the starting point until the ending point is confirmed.

(Objective of the Invention) An object of the present invention is to increase the recognition speed by detecting the beginning of the input voice and immediately starting the matching operation, and by setting the matching path in anticipation of variations in the speaking rate. The purpose is to absorb fluctuations in speaking speed.

(Summary of the Invention) The first feature of the present invention is that matching processing is performed for each frame of the input pattern, and the degree of dissimilarity corresponding to each matching pass of each standard pattern is updated and stored for each frame. It is in.

First, a method using audio power can be used to detect the presence of input audio. In this case, the start end of the audio is detected when the frame power P(j) (where j is the frame number of the input pattern) exceeds a predetermined threshold. However, due to external noise or the like, the power P(j) may exceed the threshold even if no voice input is being performed, and the starting point may be determined to be incorrect. Therefore, although the frame whose frame power P(j) exceeds the threshold is considered to be the starting point and recognition processing is started, if the frame power exceeds the threshold for three or more consecutive frames, the input frame is not considered to be the starting point of the audio. The recognition process is interrupted and the process returns to the start edge detection process. However, the frame length is set to 16 msec. Here, the numbering of the frame whose frame power exceeds the threshold from the start of the audio is defined and is called the h-th audio frame, which is distinguished from a mere input frame number. In other words, the audio frame with audio frame number h has h in the sound section.
corresponding to the input frame.

If the matching process that normalizes the speech rate can be started from the beginning of the voice and performed every cycle (frame period) in which the output of the voice analyzer is obtained, there is no need to store all the data from the voice analyzer from the start. , and the response time is also faster.

A second feature of the present invention is to set matching paths assuming cases in which utterances are performed slowly, normally, and quickly, and perform matching processing on each matching path. At the time of detecting the beginning of speech, the speech rate of the word that is about to be input is unknown. Therefore, by setting matching paths assuming that the utterance is performed slowly, normally, or quickly, and performing matching processing on each matching path, matching processing can be started even before the end is detected. becomes. Of course, in this case, there is a small possibility that there exists a path where the end of the input matches the end of the standard pattern, but the dissimilarity on the path where the end of the input most matches the end of the standard pattern is minimal. It is expected that

On the other hand, some words have a portion where the frame power is less than the threshold, such as between the "i" and "chi" in "ichi", that is, a silent frame. Such a part is called a "power dip." The length of this power dip varies depending on the word, but it usually rarely exceeds 30 frames. After detecting the start of audio, if the frame power at a certain frame time point is less than the threshold, that is, the frame becomes silent, it is difficult to determine whether that frame time point is the beginning of a power dip or the end of audio. This determination is usually made from that point onwards.
This is done depending on whether there is a frame that satisfies the audio start condition (3 or more frames in a row where the frame power is above the threshold) within 30 frames, so a decision cannot be made until 30 frames have passed at most. Therefore, the matching results when the frame power becomes less than the threshold must be suspended in some way. In the present invention, this problem is solved by stopping the updating of the audio frame number and stopping the matching process for frames whose frame power is less than a threshold value, that is, in a silent state.

FIG. 2a is a diagram showing an example of a plurality of matching passes for matching an input pattern and a standard pattern in the speech recognition method according to the present invention, FIG. 2b is a diagram showing an example of frame power of the input pattern,
Figure 2c shows the dissimilarity Dn(j), D′n(j), in each matching pass between the input pattern and the standard pattern,
FIG. 3 is a diagram showing an example of D″n(j).

In Figure 2 a, the range of speech rate is, for example, ±
The figure shows the case where three matching paths are set, assuming that the ratio is 20%. In FIG. 2a, the horizontal axis represents the frame number of the input pattern. Further, the vertical axis represents the frame number of the standard pattern, and taking the nth standard pattern Sn as an example, let its frame length be SL(n). Reference numeral 101 indicates a path assuming that the speech is uttered 20% slower, 102 a path that assumes standard speech, and 103 a path that assumes that the speech is uttered 20% faster. When the power of the j-th input frame is greater than or equal to the threshold, the distance from the standard pattern Sn on the three paths is given by the following equation. However, h
is the audio frame number corresponding to the j-th input frame number, and W(i, j) is when the input frame number is j and the channel number is i (where i = 1 to 8).
Sn (i, k) is a component of the standard pattern with frame number k and channel number i.

Distance to path 101 dn(j) = ₈ 〓 ⁱ⁼¹ | W (i, j) − Sn (i, k) | where k = [1/1.2h] [1/1.2h] ≦ SL(n) SL (n) [1/1.2h] > SL(n) ...Equation 1 Distance to path 102 d'n(j) = ₈ 〓 ⁱ⁼¹ | W (i, j) - Sn (i, k') ｜ However, k′=h SL(n) h≦SL(n) h＞SL(n) ...Second formula Distance to path 103 d″n(j)= ₈ 〓 ⁱ⁼¹ ｜W(i, j) −Sn (i, k″) | However, k″=[1/0.8h] [1/0.8h]≦SL(n) SL(k) [1/0.8h]>SL(n) ……3rd formula Note that [ ] indicates a Gaussian symbol.

According to the above equation, in path 101, the distance between the j-th input audio frame of the input pattern and the k-th frame of the standard pattern is calculated.
In pass 102, the distance between the j-th input frame of the input pattern and the k'-th frame of the standard pattern is calculated, and in the pass 103, the distance between the j-th input frame of the input pattern and the k''-th frame of the standard pattern is calculated. However, if k, k', k'' indicating the frame number of the standard pattern is larger than the length SL(n) of the standard pattern, it is limited to SL(n).

On the other hand, if the power of the j-th input frame is less than the threshold, the distances dn(j), d′n on each path
(j), d″n(j) is forced dn(j)=0...4th equation d′n(j)=0...5th equation d″n(j)=0...6th equation By using the formula, processing is performed in which dissimilarity calculations are not actually added when the frame power is less than the threshold value. For this purpose, the standard pattern is also stored in a form excluding frames corresponding to power dips, that is, frames corresponding to the silent state.

In this way, although the distance is set to 0 for frames that are not important as dissimilarities, such as power dips or matching after the end of a standard pattern, the matching is essentially linear in the present invention.

Next, the dissimilarities Dn(j), D′n(j), and D″n(j) of the input pattern up to the j-th input frame are calculated.

Dissimilarity of path 101 Dn(j)=dn(j)+Dn(j−1) ...Formula 7 Dissimilarity of path 102 D'n(j)=d'n(j)+D'n(j −1)...Equation 8 Dissimilarity of path 103 D″n(j)=d″n(j)+D″n(j−1)...Equation 9 In other words, j-th on each path To calculate the dissimilarity of frames, calculate the distance for each channel (for example, |W(i,j)−Sn(i,k)|) for each channel, and calculate the dissimilarity value for the (j-1)). These operations are performed when the j-th frame is input. When calculating the dissimilarity for the j-th input frame,
The input pattern data of the th frame, the standard pattern data corresponding to each pass, and 1
Only the frame dissimilarity data of the j-1th input frame before the frame is necessary, and the input pattern data from two or more frames before is unnecessary. Therefore, the storage area is reduced compared to the linear expansion/contraction matching method in which input patterns must be stored until the end is detected.

Figure 2c shows the dissimilarity Dn(j), D′n(j),
D″n(j), but as shown in Figure 2c, by forcing the distance value to 0 when the frame power is below the threshold, the dissimilarity Dn
(j), D′n(j), and D″n(j) are retained. Therefore, the dissimilarity at the end and the input frame 30 frames from the end (the end is detected for the first time at this point) The dissimilarities at are equal.

Next, the category is determined based on the degree of dissimilarity obtained for each standard pattern from the time when the end of the audio is detected (30 frames after the end of the audio). The input frame number at the time of end detection is j,
When the audio frame number is H, the dissimilarity of each path to the nth standard pattern is Dn(j), D′n(j),
It is given by D″n(j). There are as many sets of these dissimilarities as there are standard patterns (assumed to be N). We will describe a method for making category judgments using these dissimilarities. Judgment 1 The steps are performed as follows. First, the dissimilarities Dn(j), D′n(j), D″n(j) for each path with respect to the nth standard pattern are calculated.
One of them is selected. In this selection, L, L', and L'' given by the following equation are used for the voice frame number H at the time of voice end detection.

Path 101 L=[1/1.2H] ...Equation 10 Path 102 L'=H ...Equation 11 Path 103 L''=[1/0.8H] ...Equation 12 These values L, L', L'' is similar to the formula giving the number of frames of the standard pattern corresponding to the audio frame, but is not limited to the length of the standard pattern by SL(n). Therefore, L, L′,
L″ is unrelated to the type of standard pattern. Among these L, L′, and L″, the length of the standard pattern SL(n)
Select only the dissimilarity corresponding to the path that has the closest value to . For example, if L′ is closest to SL(n), path 102 corresponds and the dissimilarity for it is
D′n(J) is selected. DDn the selected dissimilarity
These selections can be made for each standard pattern.

Next, a second step of determination is performed. A minimum value is found for the degree of dissimilarity DDn for each standard pattern obtained in the first determination step. The category added to the standard pattern that gives this minimum value becomes the recognition result.

(Embodiment) FIG. 3 is an embodiment showing a circuit configuration for performing matching processing and determination processing in the present invention.
The operation will be explained in detail below.

In FIG. 3, 52 is an audio frame counter that counts the number of audio frames from the start edge. When the start edge is detected, the content becomes 0 by the reset pulse 50, and is input thereafter when the power of the input frame exceeds the threshold. A count up operation is performed by the count pulse 51. If the power of the input frame is less than the threshold, the count pulse 51 is not added and the output of the audio frame counter 52 is held. Let h be the audio frame number output from the audio frame counter 52. 53 is a signal representing the type of path during matching, and the number of paths is 3.
Since it is a book, it takes a value of 0 to 2. 54 is a ROM that gives frame numbers of the corresponding standard pattern on each path in the h-th audio frame. ROM
54 stores values corresponding to [1/0.8h], h, and [1/1.2h]. Let the output of the ROM 54 be l. 55 is a standard pattern number signal which gives the number of the standard pattern, and is assumed to be n. When the total number of standard patterns is N, it takes a value from 0 to N-1. 56 is n
This is a ROM that stores the length S(n) of the standard pattern for the standard pattern. 57 is the output of the ROM 54 which outputs the standard pattern frame number l, and the ROM 5 which outputs the standard pattern length SL(n).
Compare the contents of 6 and select “1” if l≦SL(n),
This is a comparator that outputs "0" if l>SL(n). 58, the output of comparator 57 is “1”
In this case, the output of ROM 54 is sent to comparator 57.
This is a selector that selects the output of the ROM 56 when the output is "0". Comparator 57 and selector 58 determine that if l≦SL(n), then l>SL(n)
If so, an operation is performed in which SL(n) is output from the selector 58. Let the output of the selector 58 be k.
59 is a signal giving channel number i. 6
0 is a standard pattern memory that is addressed by the channel number i, the output of the selector 58, for example k, and the standard pattern number signal n, and outputs each component Sn (i, k) of the standard pattern. Reference numeral 61 denotes a memory that stores one frame's worth of input pattern components W(i, j) that have undergone spectrum normalization, and is given an address by a channel number signal 59. 62 is an input terminal of the memory 61;
Input data W(i, j) whose spectrum has been normalized by a spectrum normalization unit (not shown) is input.
63 is an arithmetic unit that outputs the following values between the output W (i, j) of the memory 61 and the output Sn (i, k) of the standard pattern ROM 60 according to the control signal CONT.

When CONT = 1 | W (i, j) - Sn (i, k) | When CONT = 0 0 ...Equation 13 The CONT signal is set to "1" when the frame power is above the threshold, and when the frame power is below the threshold It is a signal that becomes "0" at this time. 64 is an adder, and 65 is an address that uses the values of the pass signal 53 and standard pattern number signal 55.
It is a RAM and stores dissimilarities Dn(j), D'n(j), and D″n(j). 67 gives the path number to be selected in the first step of judgment for the audio frame length h. However, since the pass number to be selected is given for each standard pattern, the ROM 67 inputs the audio frame number h and the standard pattern number n as addresses and outputs the pass number to be selected at that time. Output and pass signal 53
This is a comparator that outputs "1" if the two match. 69 follows the signal of the comparator 68 and when the output of the comparator 68 is "1",
This converter converts the output of the RAM 65 as it is, or converts the output of the RAM 65 into the maximum dissimilarity value when the output of the comparator 68 is "0". 70 is a minimum value selection circuit that compares the output of the converter 69 and the output of a register 71, which will be described later, and outputs the smaller value, and outputs two signals. One is a signal that gives the smaller value of the comparison results, and this signal is stored in the register 71. The other signal is a clock that is generated if the value of the converter 69 is smaller as a result of the comparison, and becomes the input clock of the register 72. register 7
1 is a register giving the minimum value of dissimilarity, and the maximum value of dissimilarity is set at the beginning of a frame period. The register 72 is a register that stores a standard pattern number signal in response to the output pulse of the minimum value selection circuit 70, and stores the number of the standard pattern that gives the minimum value of dissimilarity.

FIG. 3 is constructed as described above, and its operation will be explained below.

Each process is started when the frame power P(j) exceeds the threshold value, but the process is reset if the frame power P(j) does not exceed the threshold value for three or more consecutive frames. The counter 52 is in the reset state by the reset pulse 50 before the start frame of the audio. Also, all values in the memory 65 have been reset. Hereinafter, the processing within one frame period after the start edge detection will be sequentially explained. However, for the sake of explanation, the input frame number is assumed to be j. If the frame power of the jth input frame exceeds the threshold, a count pulse 51 is applied to the counter 52;
2 counts up and outputs the audio frame number h. The frame number of the standard pattern corresponding to the audio frame number h is outputted by the ROM 54, the ROM 56, the comparator 57, and the selector 58. The i-channel data Sn (i, k) of the k-th frame of the n-th standard pattern is stored in the ROM60.
is output by . On the other hand, input data W(i, j) after spectral normalization of the j-th input frame outputted from the previous-stage spectral normalization unit (not shown) is input to the memory 61 from the input terminal 62 and stored therein. There is. ROM60 output Sn(i,
k) and the output W(i, j) of the memory 61 are outputted in synchronization with the channel number signal 59, and the arithmetic unit 63 performs the calculation given by equation (13). Calculations corresponding to the first to ninth expressions are executed between the output of the arithmetic unit 63 and the memory 65. Actually, formulas 1 to 9
The expression operations are performed in the following unified form.

(Memory 65)←(Memory 65) +|W(i,j)−Sn(i,k″)|0 Next, the operation of the first step of determination will be explained.

ROM is used to select the path required for the first step of judgment.
67, a comparator 68, and a converter 69. The ROM 67 stores the path number to be selected in the case of audio frame number h in the nth standard pattern, and its setting standard is based on the nth path number among the calculation results of equations 10 to 12 for the audio frame number. Standard pattern length SL
given by the path number closest to (n).
The pass number output from the ROM 67 and the pass signal 53 are compared by a comparator 68, and if the two match, the comparator 68 sets "1" to the converter 6.
Output to 9. The converter 69 outputs the output of the memory 65 when the input from the comparator 68 is "1", and outputs the maximum value of dissimilarity when the input is "0". ”, that is, when the output of the ROM 67 and the path signal 53 do not match, the dissimilarity at that time is substantially prohibited from being selected by the minimum determination process. This process performs the first step of determination. Next, the minimum value selection circuit 70 stores in the register 71 the smaller of the dissimilarity stored in the register 71 and the dissimilarity output by the converter 69. At the same time, if the output of converter 69 is smaller, a pulse is applied to register 72 and the standard pattern number at that time is stored in register 72. If this process is performed for all standard patterns, the standard pattern number giving the minimum dissimilarity at that time will be stored in the register 72. The time chart for the above processing within one frame period is shown in the fourth section.
As shown in the figure.

The above processing is performed every frame period, and the result of the register 72 at the time when the end is detected becomes the final recognition result, and is output from the output terminal 73.

(Effects of the Invention) As described above, the present invention has the advantage that a recognition result can be obtained within one frame after the termination is detected because distance calculation and dissimilarity calculation are performed for each input frame. Furthermore, in order to calculate the dissimilarity, it is only necessary to accumulate the dissimilarity value for the previous input frame and the distance value for the current input frame, and there is no need to store the input data from the start to the end.

Furthermore, since it is a method aimed at simplifying the circuit configuration, it is easy to implement on an LSI, and it can be realized by providing an inexpensive LSI chip for speech recognition with a small number of gates, and at the same time by using software processing on a general-purpose microprocessor. It's something you get.

[Brief explanation of the drawing]

Figure 1 is a diagram for explaining utterance length variation, Figure 2
3 is a block diagram showing an embodiment of the present invention, and FIG. 4 is a time chart for processing within one frame period of this embodiment. FIG. 52...Counter of audio frame number h, 54
...ROM for generating something equivalent to the frame number of the standard pattern, 56...ROM storing the length of the standard pattern, 57...Comparator, 58...Selector, 60...Standard pattern memory, 61 ...Input pattern memory, 63...
Distance calculator, 64... Adder, 67... ROM of path number to be selected, 68... Comparator,
69...Converter, 70...Minimum value selection circuit,
71...Memory of minimum dissimilarity, 72...Memory of standard pattern number as recognition result.

Claims (1)

[Scope of Claims] 1. A standard pattern expressed as a frame time series of frequency components corresponding to each standard voice and with frames corresponding to a silent state removed, and a path provided corresponding to each standard pattern. (a) extracts the input pattern from the input audio as a frame time series of frequency components; (b) detects the start of the input audio and starts counting frames of the input pattern; While detecting a sound state, the audio frame number is updated sequentially, while while detecting a silent state, updating of the audio frame number is stopped, and before confirming the end of the input audio, the sound state is resumed. (c) upon each update of the audio frame number, by generating a standard pattern of frame numbers in an essentially linear relationship to that audio frame number; A plurality of matching paths are set between the pattern and each standard pattern, and (d) each time the audio frame number is updated, the distance between the input pattern and each standard pattern is calculated between the frames matched by each matching path. (e) The cumulative value of the distance along the matching path from the start of the input audio to an arbitrary audio frame number is defined as the dissimilarity, and each time the audio frame number is updated, the immediately preceding dissimilarity and the relevant frame number are calculated. (f) The dissimilarity corresponding to each matching path for each standard pattern is updated and stored by adding the distance with the distance in and storing it once, and (f) every time the audio frame number is updated, the one path selection signal is generated corresponding to each standard pattern based on the standard pattern, a dissimilarity corresponding to the matching path specified by the path selection signal is selected for each standard pattern, and among the selected dissimilarities, (g) When the end of the input audio is confirmed, the code of the standard pattern corresponding to the dissimilarity that shows the minimum value is updated and stored every time the audio frame number is updated, and (g) when the end of the input audio is confirmed, A speech recognition method characterized in that a code of the standard pattern corresponding to a correspondingly stored minimum value indicating dissimilarity is recognized as a category of input speech.